Characterization and prediction of PM2.5 levels in Afghanistan using machine learning techniques

Afghanistan faces severe air quality issues in major cities due to various sources like transportation, domestic energy use, and industrial activity. This study investigates PM2.5 spatiotemporal variability and its future relationship with six meteorological variables: precipitation, temperature, dewpoint temperature, wind speed, boundary layer height and surface pressure. This study aims to assess the spatiotemporal variability of PM2.5 concentrations in Afghanistan and derive models for predicting PM2.5 from the six variables. Satellite-measured PM2.5 and six reanalyses (ERA5) meteorological datasets for 1998–2020 were used as predictors. Three machine learning models, AdaBoost, Random Forest (RF), and Support Vector Machine (SVM), were used to develop the annual and seasonal PM2.5 concentration prediction model. Results suggest PM2.5 levels ranging from 60–80 µg/m³ in northern, southern, and western regions, while other areas experience lower levels (12–50 µg/m³). The lowest PM2.5 concentrations are in the Hindu Kush mountain range. Summer exhibited the highest PM2.5 concentrations, reaching a maximum of 137.4 µg/m³ and an average of 48.5 µg/m³. Among the prediction models, RF performed best in predicting PM2.5 across Afghanistan, as evidenced by the evaluation metrics: NRMSE (59.2), RSR (0.59), rSD (0.75), and higher values of NSE (0.65), R² (0.65), and KGE (0.68). The geographical and seasonal distribution of observed PM2.5 distribution was very similar to the PM2.5 estimated using RF compared to the other two models. The analysis showed that air temperature, precipitation, wind speeds, and boundary layer heights play significant roles in PM2.5 distribution. However, the relationship between precipitation and PM2.5 was more pronounced than other meteorological variables.